Which open generative AI models are most suitable for a speech-to-text transcription task in an LLM-based application where speed is a critical requirement?